Large model-based orbiting robot positioning error compensation method

Abstract

The application discloses a large model-based track inspection robot positioning error compensation method and relates to the technical field of robot positioning, and comprises the following steps: extracting spatial curvature features in a track image and inertial signals to generate a track change perception sequence; separating a stable section and a curvature mutation section based on end data of the sequence, and differentiating and mapping to obtain a track inertia deviation signal; establishing an adaptive weight according to the inertia deviation signal and generating a curvature response instruction; fusing environment recognition features and track sleeper positioning data to form a posture correction instruction stream; and combining forward visual information to generate a path inertia suppression output and complete a travel path adjustment. Through track change perception and inertia deviation cooperative processing, the application realizes continuous perception and dynamic response to track curvature change, weakens track inertia dependence, and improves positioning stability; and through prospective posture and path adjustment, the application guarantees continuous travel and stable posture in a complex bending section and enhances the safety and reliability of inspection operation.

Description

Technical Field

[0001] This invention relates to the field of robot positioning technology, and more specifically to a method for compensating positioning errors of a track-following robot based on a large model. Background Technology

[0002] Track-inspection robots are intelligent mobile platforms used in rail transit, smart manufacturing, and industrial inspection to automatically perform track inspection, equipment maintenance, and environmental monitoring tasks. They typically operate along tracks, equipped with lidar, vision cameras, infrared sensors, inertial measurement units, and GPS / BeiDou positioning modules to collect and identify trajectory data, equipment status, and environmental parameters in real time, thereby replacing manual inspections and improving efficiency and safety. However, in tunnels, stations, or complex metallic environments, positioning signals are easily obstructed and interfered with, leading to problems such as position drift, path deviation, and unstable repositioning for the track-inspection robot.

[0003] To address this core pain point, a large-model-based method for compensating positioning errors in track-following robots has emerged. This method utilizes a large industrial model with cross-modal understanding and adaptive learning capabilities to fuse heterogeneous data from multiple sensors, such as visual features, inertial data, magnetic field information, and GNSS signal residuals. Through deep feature mapping and temporal correlation modeling, it automatically identifies the spatial distribution and temporal evolution of positioning errors. Unlike traditional filtering or few-sample regression, the large model can not only transfer and learn error patterns across different scenarios but also achieve online dynamic compensation during trajectory execution. This allows the robot to maintain high-precision positioning and path consistency even under complex track conditions, thereby significantly improving the reliability, autonomy, and environmental adaptability of inspection tasks.

[0004] The existing technology has the following shortcomings:

[0005] When using large models for positioning error compensation of track-following robots, the models, trained under fixed trajectory characteristics for extended periods, can easily become overly reliant on existing spatial patterns. When sudden changes in spatial curvature occur in the actual track environment, the model will still output predictions based on the original trajectory trend, leading to delayed path judgments or even directional reversals. This trajectory reversal phenomenon can cause the robot to turn around or enter restricted areas, affecting inspection accuracy and potentially causing equipment collisions, communication interruptions, and operational disruptions.

[0006] The information disclosed in the background section is only intended to enhance the understanding of the background of this disclosure, and therefore may include information that does not constitute prior art known to those skilled in the art. Summary of the Invention

[0007] The purpose of this invention is to provide a positioning error compensation method for a track-following robot based on a large model, so as to solve the problems in the background art mentioned above.

[0008] To achieve the above objectives, the present invention provides the following technical solution: a positioning error compensation method for a track-following robot based on a large model, comprising the following steps:

[0009] Step 1: Extract the spatial curvature features from the trajectory image data and inertial signal data collected during the operation of the track-following robot. Generate a trajectory change perception sequence based on the extracted spatial curvature features. Capture the dynamic bending trend of the track in the trajectory change perception sequence and mark the curvature abrupt change nodes that exist during the trajectory change process to provide time-continuous basic data for subsequent trajectory adjustment.

[0010] Step 2: Based on the time window data consisting of multiple consecutive time points selected backward from the current timestamp in the trajectory change perception sequence, the continuity characteristics of the stable track segment and the curvature change segment are dynamically separated. The historical running trend of the track is differentially mapped with the real-time trajectory offset data to obtain the trajectory inertial offset signal, and this trajectory inertial offset signal is used as the offset reference for direction correction.

[0011] Step 3: Based on the trajectory inertial offset signal, establish an adaptive weight allocation relationship for curvature abrupt change nodes, dynamically balance the inertial continuation trend and the real-time offset trend, and generate curvature response commands based on the balance results to weaken the trajectory reversal delay effect caused by inertial continuation.

[0012] Step 4: Based on the curvature response command, the environmental recognition features contained in the trajectory image data are fused with the spatial positioning data between the sleepers to form a local attitude correction command stream. The trajectory of the track-following robot is continuously adjusted according to the local attitude correction command stream, so as to achieve a smooth transition of the trajectory turning segment and avoid the occurrence of direction reversal.

[0013] Step 5: Based on the local attitude correction command stream, the visual dynamic distribution information in front of the trajectory is fused to generate path inertia suppression output. Based on the path inertia suppression output, the travel path of the track-following robot is proactively adjusted in the curvature change region to ensure the travel continuity and attitude stability of the track-following robot in complex curves.

[0014] Preferably, the step of generating a trajectory change sensing sequence based on the extracted spatial curvature features includes:

[0015] As the track-following robot moves along the track, it continuously acquires trajectory image data through multi-angle vision acquisition devices installed at the front and bottom of the vehicle body, and records the images in time sequence through timestamps. At the same time, the inertial measurement unit synchronously outputs triaxial acceleration, angular velocity and attitude angle information at the same time intervals, forming an inertial signal data sequence that is time-aligned with the trajectory image data.

[0016] After obtaining time-consistent trajectory image data and inertial signal data, the spatial morphology information of the track is extracted based on the geometric structural features in the trajectory image data. The angular velocity change and linear acceleration direction in the inertial signal data are matched with the track curvature region in the track image data to generate a set of spatial curvature features containing the trajectory position, direction change and attitude curvature.

[0017] Based on the set of spatial curvature features, with the direction of travel of the track-following robot as the main axis, the spatial curvature features corresponding to each moment are arranged in chronological order and a continuous correlation is established to generate a trajectory change perception sequence, which is used to reflect the dynamic bending trend of the track.

[0018] Based on the curvature change relationship at each time point in the trajectory change sensing sequence, the rate of change of trajectory curvature is calculated. When the rate of change exceeds the set change threshold, the corresponding spatial position is identified as a curvature mutation node, and a mutation node chain is established to provide time-continuous basic data for trajectory adjustment.

[0019] Preferably, the step of differentially mapping the historical trajectory trend with real-time trajectory offset data to obtain the trajectory inertial offset signal includes:

[0020] During track operation, the track-following robot reads the trajectory space curvature value, attitude angle, direction vector and track displacement information frame by frame from the time window data interval consisting of multiple time points continuously selected forward from the current timestamp in the trajectory change perception sequence. It also performs spatial reconstruction by combining the sleeper, track edge line and center line features in the trajectory image data. At the same time, it matches the angular velocity change, linear acceleration direction and attitude angle change amplitude in the inertial signal data with the trajectory curvature change to build a structured trajectory operation dataset.

[0021] After obtaining the structured trajectory running dataset, the magnitude and direction of the change in the trajectory space curvature value are analyzed. The time series data of the time window data interval, which is composed of multiple time points continuously selected forward from the current timestamp as the endpoint, is separated by continuous feature separation. The part of the trajectory space curvature value change magnitude continuously less than the magnitude threshold is classified into the stable segment set, and the part of the trajectory curvature change rate exceeds the change threshold is classified into the abrupt change segment set. Smoothing processing is performed in the boundary area between the track stable segment and the curvature abrupt change segment to form the trajectory segment structure.

[0022] Based on the feature data in the stable segment set and the abrupt segment set, differential mapping is performed on the difference between trajectory direction change, attitude deflection and orbit centerline displacement according to the time correspondence, and a trajectory inertial offset signal containing the trajectory inertial continuity direction, offset amplitude and duration is generated as an offset reference for the orientation correction of the track-following robot.

[0023] Preferably, during the generation of the trajectory inertial offset signal, the differential mapping of the trajectory direction change, attitude deflection, and orbit centerline displacement difference is continuously calculated in time sequence, and the three parameters of inertial continuation direction, offset amplitude, and duration are recorded in the trajectory inertial offset signal to ensure that the trajectory inertial offset signal maintains temporal continuity and spatial correspondence during the operation of the track-following robot.

[0024] Preferably, the step of generating curvature response instructions based on the equilibrium result includes:

[0025] Based on the trajectory inertial offset signal, spatial location identification and temporal correlation are performed on the curvature change nodes. The trajectory inertial continuation direction, offset amplitude, duration and attitude angle change are extracted and matched with the sleeper spacing change, track edge line turning angle and track center line bending radius in the trajectory image data to construct a node comprehensive dataset containing inertial offset features and curvature change features.

[0026] Based on the node comprehensive dataset, an adaptive weight allocation relationship is established for each curvature change node. According to the correspondence between the trajectory inertial offset signal and the real-time offset trend in the two dimensions of time and space, the ratio of the inertial continuation trend weight to the real-time offset trend weight is adjusted to generate a dynamically balanced weight allocation result.

[0027] Based on the weight allocation results and trajectory inertial offset signal, combined with the spatial geometric features and attitude change features of curvature change nodes, the adjustment direction, adjustment amplitude and action time of curvature response commands are determined, and curvature response commands are generated in a continuous output manner to weaken the trajectory reversal delay effect caused by inertial continuation.

[0028] Preferably, the adjustment direction of the curvature response command is determined by the vector superposition of the inertial continuation trend direction after dynamic equilibrium and the real-time offset trend direction. The adjustment range is set according to the offset range of the trajectory inertial offset signal, and the action time is determined according to the duration of the trajectory inertial offset signal. The curvature response command is output before the track-following robot enters the curvature change region, so as to realize the early correction of the trajectory direction and the continuous adjustment of the attitude.

[0029] Preferably, the step of fusing environmental recognition features contained in the trajectory image data with spatial positioning data between sleepers to form a local attitude correction command stream includes:

[0030] After obtaining the curvature response command, the spatial region corresponding to the curvature response command is extracted from the trajectory image data. The geometry, arrangement direction, sleeper spacing, spatial orientation of the track centerline, curvature trend of the track edge line, and boundary features on both sides of the track are identified. Based on the time index, the environmental identification features are synchronized with the curvature response command in time and space.

[0031] Based on the spatial position parameters of the curvature response command, the environmental identification features are fused with the spatial positioning data between sleepers. By matching the coordinates of the track centerline, comparing the sleeper spacing, and associating the normal angle of the sleeper surface, a comprehensive trajectory spatial structure description is formed, which includes the track geometry, sleeper distribution pattern, and attitude inclination angle.

[0032] Combined with the adjustment direction of the curvature response command, the trajectory adjustment direction is identified in the fused integrated trajectory space structure. Based on the sleeper spacing, track centerline direction, track edge line turning angle and track tilt angle, a continuous local attitude correction command flow is generated. The local attitude correction command flow includes trajectory adjustment direction, attitude correction angle, adjustment magnitude and adjustment duration.

[0033] Based on the generated local attitude correction command stream, the trajectory of the track-following robot is continuously adjusted. By controlling the forward direction, attitude angle, turning rate and adjustment time, a smooth transition of the trajectory turning segment is achieved and the phenomenon of direction reversal is avoided.

[0034] Preferably, during the continuous adjustment of the trajectory of the track-following robot based on the generated local attitude correction command stream, the trajectory adjustment includes continuously correcting the forward direction within the range of curvature response command, and updating the attitude correction angle in real time when the sleeper spacing changes synchronously with the turning angle of the track edge line. By controlling the adjustment amplitude and adjustment duration, the spatial consistency between the trajectory direction and the track centerline direction is achieved.

[0035] Preferably, the steps of generating path inertia suppression output by fusing visual dynamic distribution information ahead of the trajectory based on the local attitude correction command stream include:

[0036] Based on the local attitude correction command stream, the visual dynamic distribution information in front of the track is collected and identified. According to the trajectory adjustment direction and range of action in the local attitude correction command stream, the visual acquisition device acquires an image sequence containing sleepers, track center line, track edge line, tunnel inner wall and trackside obstacles. The visual dynamic distribution information is synchronized with the attitude data according to the time index to obtain the spatial geometric features in front of the track.

[0037] After acquiring the visual dynamic distribution information in front of the track, it is fused with the local attitude correction command stream. Based on the time correspondence between the visual dynamic distribution information and the local attitude correction command stream, the track adjustment direction is matched with the track centerline direction, the curvature of the track edge line is matched with the attitude correction angle, and the sleeper spacing change is compared with the adjustment range to form a comprehensive dataset containing track attitude and spatial geometric features.

[0038] After obtaining the integrated trajectory attitude and visual distribution data, the path inertia suppression output, which includes the path adjustment direction, inertia suppression amplitude and adjustment duration, is generated based on the curvature direction, curvature change rate and attitude correction angle of the track centerline. The path of the track-following robot is then proactively adjusted in the curvature change region through continuous output to maintain the continuity of movement and attitude stability.

[0039] Preferably, during the generation process, the path inertia suppression output determines the path adjustment direction based on the bending direction and curvature change rate of the track centerline, and determines the inertia suppression amplitude based on the attitude correction angle in the local attitude correction command stream and the sleeper spacing change in the visual dynamic distribution information. By controlling the look-ahead adjustment duration, the path inertia suppression output is made to act in advance before the track-following robot enters the curvature change region and gradually weaken after passing through the curvature change region.

[0040] The technical effects and advantages provided by the present invention in the above technical solution are as follows:

[0041] This invention introduces a collaborative processing mechanism combining trajectory change sensing sequences and trajectory inertial offset signals, enabling a track-following robot to continuously perceive dynamic changes in the curvature of the track space during operation and continuously characterize the trajectory trend before and after curvature abrupt changes. By distinguishing between stable track segments and curvature abrupt change segments, and generating curvature response commands based on this distinction, the model's inertial dependence on existing trajectory trends is effectively reduced, allowing the positioning error compensation process to adjust in real time according to changes in track geometry. Therefore, the track-following robot can maintain the continuity of path judgment in complex track environments, reducing trajectory reversals and directional deviations caused by prediction lag, and improving the overall stability and reliability of positioning results.

[0042] Guided by curvature response commands, this invention integrates environmental recognition features, spatial positioning data between sleepers, and visual dynamic distribution information ahead of the track to construct a proactive adjustment process that flows from local attitude correction commands to path inertia suppression output. This process enables the track-inspecting robot to gradually correct its attitude and direction of travel before entering curves or areas of sudden curvature changes, thus achieving a smooth transition within turning sections. This proactive path adjustment method avoids problems such as direction reversal and discontinuous travel, enhancing the continuity of travel and attitude stability of the track-inspecting robot under complex curvature conditions, and providing a safer and more reliable operational guarantee for long-term autonomous inspection operations. Attached Figure Description

[0043] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0044] Figure 1 This is a flowchart of the positioning error compensation method for a track-following robot based on a large model, as described in this invention.

[0045] Figure 2 This is a flowchart illustrating the generation of a trajectory change sensing sequence based on the extracted spatial curvature features in this invention.

[0046] Figure 3 This is a flowchart illustrating the process of obtaining the trajectory inertial offset signal according to the present invention;

[0047] Figure 4 This is a flowchart illustrating the generation of curvature response commands based on the equilibrium results of this invention. Detailed Implementation

[0048] Exemplary embodiments will now be described more fully with reference to the accompanying drawings. However, these exemplary embodiments can be implemented in many forms and should not be construed as limited to the examples set forth herein; rather, they are provided so that the description of this disclosure will be more complete and fully convey the concept of the exemplary embodiments to those skilled in the art.

[0049] This invention provides, for example Figures 1 to 4 The positioning error compensation method for a track-following robot based on a large model, as shown, includes the following steps:

[0050] Step 1: Extract the spatial curvature features from the trajectory image data and inertial signal data collected during the operation of the track-following robot. Generate a trajectory change perception sequence based on the extracted spatial curvature features. Capture the dynamic bending trend of the track in the trajectory change perception sequence and mark the curvature abrupt change nodes that exist during the trajectory change process to provide time-continuous basic data for subsequent trajectory adjustment.

[0051] The specific steps for generating a trajectory change sensing sequence based on the extracted spatial curvature features are as follows:

[0052] As the track-following robot moves along the track, it continuously acquires trajectory image data through multi-angle vision acquisition devices installed at the front and bottom of the vehicle body. The acquired images cover the sleepers, track surface, track edge lines, tunnel walls, and environmental areas on both sides of the track. These vision acquisition devices continuously capture image frames at a fixed frame rate and record them using timestamps to achieve temporal arrangement of the images. Simultaneously, the inertial measurement unit inside the track-following robot synchronously outputs triaxial acceleration, angular velocity, and attitude angle information at equal time intervals, thus forming a corresponding inertial signal data sequence. To maintain data synchronization, the trajectory image data and inertial signal data are recorded and correlated according to a unified time base, and all timestamps correspond to the system clock sequence, ensuring precise time alignment between image frames and inertial signals at the same moment. In this way, continuous trajectory image sequences and inertial signal sequences can be obtained simultaneously in both spatial and temporal dimensions. Subsequently, the original trajectory image sequence is preprocessed to remove blurred, ghosted, and unevenly lit image frames to ensure clear contours of the track edge lines and sleeper structures, thus providing high-quality image data input for subsequent spatial curvature feature extraction.

[0053] After obtaining time-consistent trajectory image data and inertial signal data, the spatial morphology information of the track is extracted based on the geometric structural features in the trajectory image data. By analyzing the spatial positional changes of the track centerline, sleeper spacing, track edges, and tunnel walls frame by frame, the geometric orientation and curvature of the trajectory in three-dimensional space are extracted. For each frame, the spatial extension direction of the sleepers and track edges is determined, and their angular changes relative to the robot's forward direction are recorded. Subsequently, the inertial signal data at the corresponding time is matched, and the changes in angular velocity and linear acceleration are correlated with the curved areas of the track in the trajectory image data, thus reflecting the dynamic attitude changes of the trajectory in the time dimension. Through this matching method, the spatial morphology information of the trajectory and the inertial change features form a one-to-one correspondence in space and time, generating a spatial curvature feature set containing the trajectory position, direction changes, and attitude curvature degree. This spatial curvature feature set records the geometric changes of the track at different positions and times, and can comprehensively describe the spatial orientation and curvature trend of the track-following robot's running path.

[0054] To facilitate understanding of the technical solution of the present invention, the above implementation process is illustrated below with examples:

[0055] The track-following robot collects trajectory image data at a frequency of 30 frames per second while running along the track, and the inertial measurement unit (IMU) outputs inertial signal data at a frequency of 100Hz. Assume that within a continuous time period from t1 to t4, the robot's front-end camera sequentially acquires four track images, with the sleeper projection spacing in the images being 28 pixels, 26 pixels, 24 pixels, and 22 pixels, respectively. The angle between the two edges of the track gradually changes from 0° to 12°. During this time, the IMU records angular velocity changes of 0.05 rad / s, 0.08 rad / s, 0.11 rad / s, and 0.13 rad / s, corresponding to linear accelerations in the longitudinal direction of 0.3 m / s², 0.35 m / s², 0.4 m / s², and 0.45 m / s², respectively. By correlating the image frames with the inertial signals using timestamps, the spatial relationship of the track centerline can be obtained: the decrease in sleeper spacing reflects the beginning of track curvature, and the increase in the angle between the two edges of the track corresponds to a rightward turning trend. By aligning the angular velocity increment with the position of the curved region in the image, the curvature direction and intensity of the track centerline within that time period are determined. The track centerline position, track edge angle, and inertial angular velocity data at each moment are jointly recorded to form a set of spatial curvature feature data. For example, data is recorded at time t1 (track angle 0°, angular velocity 0.05 rad / s), at time t2 (angle 4°, angular velocity 0.08 rad / s), at time t3 (angle 8°, angular velocity 0.11 rad / s), and at time t4 (angle 12°, angular velocity 0.13 rad / s). Thus, a one-to-one correspondence is established between the track geometric changes in the trajectory image data and the attitude changes in the inertial signal data over time, generating a set of spatial curvature features containing changes in track position, direction, and attitude curvature.

[0056] Subsequently, a trajectory change perception sequence is generated based on the extracted set of spatial curvature features. Using the robot's direction of travel as the main axis of the sequence, the spatial curvature features corresponding to each moment are arranged chronologically to form a temporal curvature chain reflecting the dynamic changes in the trajectory. During the establishment of this trajectory change perception sequence, the spatial curvature features at adjacent time points are continuously correlated, forming a dynamic description of the trajectory curvature changing over time. Through this arrangement and correlation, the trajectory change perception sequence not only includes the spatial curvature of the trajectory but also its temporal variation. This trajectory change perception sequence is continuously updated during the robot's operation; with each new spatial curvature feature at a new time point, the sequence automatically expands, forming a complete curvature change transition between consecutive frames. In this way, the trajectory change perception sequence can continuously reflect the straight sections, turning sections, and curvature change regions of the track, enabling the robot to maintain a continuous perception of the dynamic geometric changes in the trajectory. Through this trajectory change perception sequence, the geometric state of the track is encoded in a temporally continuous manner into a data foundation that can be used to determine curvature trends, providing a continuous and stable input for subsequent curvature abrupt change recognition.

[0057] After the trajectory change sensing sequence is established, curvature abrupt change nodes are identified based on the curvature change relationship at each time point in the sequence. The rate of change of trajectory curvature is calculated by continuously reading the curvature changes at adjacent time points in the trajectory change sensing sequence. When the rate of curvature change at a certain time point exceeds a set threshold, the spatial location of the trajectory corresponding to that time point is identified as a curvature abrupt change node. The location of the curvature abrupt change node is spatially mapped to the track geometric features in the trajectory image data to obtain its specific coordinates in the track coordinate system. Simultaneously, each curvature abrupt change node is sequentially associated with the preceding and following time points in the trajectory change sensing sequence, forming a chain of abrupt change nodes arranged chronologically. This chain reflects the abrupt distribution of trajectory curvature changes over time, enabling the tracking robot to obtain trend information on track geometric changes before the trajectory enters a turning area or curved section. In this way, curvature abrupt change nodes not only provide the location basis for track spatial geometric changes but also form a continuous temporal information chain connected to the trajectory change sensing sequence, providing accurate basic data for subsequent trajectory adjustments and attitude corrections. The tracking robot can adjust its motion posture and turning angle in advance based on these nodes, so that the trajectory change process remains continuous and stable in the curvature change region, avoiding path deviation or direction reversal caused by trajectory geometric change.

[0058] Through the continuous execution of the above steps, spatial curvature features are extracted from trajectory image data and inertial signal data, trajectory change perception sequences are generated, dynamic bending trends are captured, and curvature abrupt change nodes are identified. The entire process achieves synchronous description of track morphology changes in both spatial and temporal dimensions, ensuring that the track-following robot can continuously monitor changes in the track's geometric features during operation. The identification results of curvature abrupt change nodes directly serve as the input basis for subsequent trajectory inertial offset signal extraction, adaptive weight balancing, and attitude correction, enabling the track-following robot to achieve continuous and stable path adjustment and precise positioning in complex curved environments.

[0059] Step 2: Based on the time window data consisting of multiple consecutive time points selected backward from the current timestamp in the trajectory change perception sequence, the continuity characteristics of the stable track segment and the curvature change segment are dynamically separated. The historical running trend of the track is differentially mapped with the real-time trajectory offset data to obtain the trajectory inertial offset signal, and this trajectory inertial offset signal is used as the offset reference for direction correction.

[0060] The specific steps for obtaining the trajectory inertial offset signal by differential mapping between the historical trajectory trend and real-time trajectory offset data are as follows:

[0061] As the tracking robot operates on the track, it continuously generates a trajectory change perception sequence. This sequence contains spatial curvature, attitude changes, and orientation information of the trajectory at consecutive time points. To obtain real-time characteristics of trajectory changes, the data within a time window segment of this trajectory change perception sequence, selected consecutively from the current timestamp, is read frame-by-frame and structured. Specifically, a data segment of several time points is selected from the end of the trajectory change perception sequence, and the trajectory spatial curvature value, attitude angle, orientation vector, and track displacement information at each time point are integrated to form a curvature change dataset within the time window. Subsequently, based on the sleeper, track edge line, and centerline information extracted from the trajectory image data, the track geometry within the same time window is spatially reconstructed to obtain the local curvature shape of the trajectory. Simultaneously, the angular velocity changes, linear acceleration directions, and attitude angle changes at corresponding time points in the inertial signal data are matched with the trajectory curvature changes to obtain the dynamic attitude change relationship of the trajectory in the time dimension. Throughout this process, all data is kept time-synchronized, and the complete correspondence of data within the time window segment of the trajectory change perception sequence is achieved through timestamp alignment. After processing, the time window data intervals in the trajectory change perception sequence, which are selected consecutively forward from the current timestamp, are constructed into a structured trajectory running dataset. This structured trajectory running dataset not only preserves the spatial shape of the trajectory, but also includes the characteristics of attitude change and motion inertia.

[0062] After obtaining a structured trajectory operation dataset, the stable segments and curvature abrupt change segments of the track are separated by continuity features. Specifically, the rate and direction of change of the spatial curvature value of the trajectory are analyzed point by point from the beginning to the end of the trajectory change sensing sequence. When the change amplitude of the spatial curvature value of the trajectory is continuously less than the amplitude threshold at several consecutive time points and the track direction vector remains parallel, it indicates that the trajectory is in the stable segment. At this time, the track geometry is continuous, the attitude direction changes slowly, and the amplitudes of angular velocity and linear acceleration in the inertial signal data remain constant. For these time points, their spatial curvature change information is uniformly classified into the stable segment set to represent the smooth operation state of the trajectory during this period. When the rate of curvature change exceeds a preset change threshold, the deflection angle of the track direction vector exceeds a set direction threshold, or the cumulative change amplitude of the attitude angle exceeds the attitude threshold over a continuous time period, it indicates that the trajectory has undergone a spatial geometric abrupt change, indicating that the trajectory is in the curvature abrupt change segment. The data within this time interval is extracted separately, and a correspondence between time index and spatial position is established to form a set of abrupt change segments. In this way, the temporal data within the time window of the trajectory change sensing sequence is separated into two parts: a stable orbit segment and a curvature abrupt change segment. To ensure the continuity of the distinction, a smooth transition is performed at the boundary between the stable orbit segment and the curvature abrupt change segment, so that the curvature change gradients on both sides of the boundary remain spatially consistent. Ultimately, the time window of the trajectory change sensing sequence is divided into several characteristic segments with clear spatial boundaries and time spans. Each segment has its own continuity characteristics, thus constructing the changing structure of the orbital geometry within a short time range.

[0063] It should be noted that:

[0064] Amplitude threshold, variation threshold, direction threshold, and attitude threshold are all judgment parameters used to distinguish between stable track segments and curvature abrupt change segments. Essentially, they are reference boundary values ​​for quantifying and delineating the degree of trajectory change. Specifically, the amplitude threshold limits the allowable range of changes in the spatial curvature value of the trajectory between adjacent time points. When the change in spatial curvature value at several consecutive time points is within this amplitude threshold range, it indicates that the track geometry maintains a continuous change state within that time interval. The variation threshold limits the boundary of the rate of change of spatial curvature of the trajectory per unit time. When the rate of change of curvature exceeds this variation threshold, it indicates that the degree of track curvature undergoes an abrupt change in the time dimension. The direction threshold limits the range of deflection angles of the track direction vector within a continuous time period. When the deflection angle of the direction vector exceeds this direction threshold, it indicates that the track's direction of travel has undergone a significant change. The attitude threshold limits the cumulative variation of the robot's attitude angles within a continuous time period. When the cumulative variation of attitude angles exceeds this attitude threshold, it indicates that the robot's attitude adjustment amplitude and the change in track curvature exhibit a synchronous abrupt change trend.

[0065] The above thresholds are preset or adaptively set based on track type, running speed and sensor resolution accuracy, and work together as a judgment standard within the time window data interval of the trajectory change sensing sequence to achieve the separation of the continuity characteristics of track stable segments and curvature change segments.

[0066] After dynamically separating the stable track segment from the curvature abrupt change segment, the separated feature data is used as input to calculate the differential mapping between the historical track trend and the real-time trajectory offset data, thereby obtaining the trajectory inertial offset signal. This trajectory inertial offset signal is then used as the offset reference for direction correction. In the specific implementation, the latest set of time points in the stable segment set is first used as the historical track trend reference. This data includes continuous direction information, attitude continuity, and spatial location characteristics of the curvature-stable segment before the current time point. Then, real-time trajectory offset data adjacent to the historical track trend time is selected from the abrupt change segment set, and its spatial direction vector, attitude change angle, and track centerline position offset are extracted. The two are then mapped one-to-one according to time sequence, and the differences in trajectory direction change, attitude deflection, and track centerline displacement are compared to generate the differential result of the trajectory in spatial geometry. This differential result reflects the offset characteristics of the trajectory under a stable operating trend due to curvature abrupt changes, and can intuitively represent the difference between the trajectory inertial continuity trend and the real-time offset trend. To quantify this difference result into a reference signal that can be used for subsequent orientation correction, the difference results are time-series integrated to obtain a continuous trajectory inertial offset signal that includes changes in time, spatial position, and attitude. This signal consists of three parts: first, the trajectory inertial continuity direction, which characterizes the motion trend of the trajectory under inertial action; second, the offset amplitude, which reflects the degree of displacement of the trajectory's spatial position relative to the ideal trajectory center; and third, the duration, which describes the temporal extension of the inertial offset. The trajectory inertial offset signal is continuously updated during the operation of the track-following robot. Whenever new end data is added to the trajectory change sensing sequence, the system recalculates and generates a new trajectory inertial offset signal, enabling the track-following robot to monitor the trajectory's inertial offset in space in real time.

[0067] Through the implementation of the above steps, the end data based on the trajectory change sensing sequence achieves dynamic separation of the continuity characteristics of the stable track segment and the curvature abrupt change segment, and realizes differential mapping between the historical track operation trend and real-time trajectory offset data, ultimately generating a trajectory inertial offset signal. This trajectory inertial offset signal accurately reflects the dynamic relationship between the inertial continuity trend and the real-time offset trend of the track-following robot in the curvature abrupt change region, providing a precise offset reference for orientation correction. Through this implementation method, the track-following robot can perceive the formation process of inertial offset in the early stages of track geometry change and adjust its attitude according to the trajectory inertial offset signal, thereby maintaining path continuity, orientation stability, and operational safety in complex curvature track environments.

[0068] Step 3: Based on the trajectory inertial offset signal, establish an adaptive weight allocation relationship for curvature abrupt change nodes, dynamically balance the inertial continuation trend and the real-time offset trend, and generate curvature response commands based on the balance results to weaken the trajectory reversal delay effect caused by inertial continuation.

[0069] The specific steps for generating curvature response commands based on the equilibrium results are as follows:

[0070] Based on trajectory inertial offset signals, spatial location identification and temporal correlation are performed on curvature abrupt change nodes to establish a comprehensive node dataset. During operation, the tracking robot continuously captures spatial curvature changes in its trajectory through a trajectory change sensing sequence. When the rate of curvature change exceeds a preset threshold or the amount of trajectory direction change exceeds a set directional offset threshold, the region is identified as a curvature abrupt change node. For each curvature abrupt change node, trajectory inertial offset signals are extracted from its immediate and adjacent time periods to form a local inertial data set. This local inertial data set includes the trajectory inertial continuation direction, offset amplitude, duration, and trajectory attitude change angle. The trajectory inertial continuation direction reflects the robot's forward trend under inertial influence; the offset amplitude represents the spatial distance of the trajectory deviating from the centerline; the duration reflects the time span of inertial action; and the trajectory attitude change angle describes the continuous change in the robot's attitude during turning. During data extraction, a time index is used to map the trajectory inertial offset signals to the timestamps of the curvature abrupt change nodes, ensuring that the local inertial data set is synchronized with the temporal position of the curvature abrupt change nodes. Simultaneously, the spatial locations of curvature abrupt change nodes are spatially paired with the track geometry in the trajectory image data to extract sleeper spacing changes, track edge turning angles, and track centerline curvature radii. These data are then matched with inertial offset information for the corresponding time periods to construct a comprehensive node dataset containing both inertial offset and curvature change features, providing a complete source of input data for subsequently establishing adaptive weight allocation relationships.

[0071] Based on the node-integrated dataset, an adaptive weight allocation relationship is established for each curvature abrupt change node to dynamically balance the inertial continuation trend and the real-time offset trend. The adaptive weight allocation relationship achieves balance by adjusting the ratio of the inertial trend weight to the real-time trend weight in both time and space dimensions. Specifically, firstly, the inertial continuation direction, offset amplitude, and duration of the trajectory inertial offset signal in the node-integrated dataset are read, and these parameters are used as input for the inertial continuation trend. Then, the real-time offset trend data at the curvature abrupt change node, including the trajectory direction deflection angle, curvature change rate, orbital centerline deviation, and robot attitude angle change rate, are read, and used as input for the real-time offset trend. Based on the time index correspondence between the two types of data, the change ratio of the inertial continuation trend and the real-time offset trend is calculated within the same time range, thus forming the initial weight allocation relationship. After the initial weight allocation is completed, the weight ratio is dynamically adjusted according to the duration of the trajectory inertial offset signal and the spatial change amplitude of the curvature abrupt change node. When the trajectory inertial offset signal is temporally extended and has a large offset amplitude, the weight of the real-time offset trend is increased to enhance the robot's response to real-time trajectory changes. When the rate of curvature change at curvature abrupt nodes remains continuous and the change in trajectory direction deflection angle is less than the direction change threshold, the weight of the inertial continuation trend is increased to maintain the smoothness of trajectory motion. Each weight adjustment is updated synchronously with the spatial curvature change of the node and the temporal change of the trajectory inertial offset signal, ensuring that the inertial continuation trend and the real-time offset trend remain continuously coordinated during the change process. Through this dynamic balancing method, the adaptive weight allocation relationship can be adjusted in real time under different curvature conditions, enabling the tracking robot to maintain consistency between trajectory direction and attitude in various trajectory change scenarios.

[0072] For example, when the tracking robot traverses a curved section of the track, the trajectory inertial offset signal recorded in the node integrated data set shows that the trajectory inertial continuation direction is 5 degrees east of north of the forward direction, with an offset amplitude of 0.12 meters and a duration of 2.4 seconds. At this time, the real-time offset trend information is read from the data corresponding to the curvature change node. The trajectory direction deflection angle is 22 degrees east of north, the curvature change rate is 0.08 radians per second, the deviation of the track centerline is 0.15 meters, and the robot attitude angle change rate is 3.5 degrees per second. Based on the time index correspondence between the two sets of data, the directional offset and positional offset of the inertial continuation trend and the real-time offset trend are compared within the same 2.4-second time range. The changes in the inertial continuation direction and the real-time direction are matched according to numerical proportions, and the ratio of their changes within the same time period is calculated. This ratio is used as the basis for the initial weight allocation calculation, thereby generating a preliminary weight allocation relationship between the inertial continuation trend and the real-time offset trend on the time series.

[0073] After achieving dynamic equilibrium of the adaptive weight allocation relationship, curvature response commands are generated based on the equilibrium results to mitigate the trajectory reversal delay effect caused by inertia continuation. The generation of curvature response commands uses the dynamically balanced weight ratios and trajectory inertial offset signals as core inputs, combined with the spatial geometric features and attitude change characteristics of curvature abrupt change nodes, to output a set of commands for trajectory direction correction. In specific implementation, firstly, the reference direction for trajectory direction adjustment is determined based on the dynamic equilibrium results, with trends with larger weight proportions used as the primary direction and trends with smaller weight proportions used as auxiliary directions. The effective range of the curvature response commands is determined based on the spatial location of the curvature abrupt change nodes and the turning angle of the track centerline, ensuring that the commands cover the trajectory areas before and after the curvature abrupt change nodes. Subsequently, the inertial continuation trend direction obtained from the equilibrium results is vector-superimposed with the real-time offset trend direction to form a unified curvature response direction. Then, based on the offset amplitude and duration of the trajectory inertial offset signal, the adjustment amplitude and duration of the curvature response commands are determined, ensuring that the adjustment process is synchronized with the dynamic changes in the trajectory. The curvature response commands are sent as continuous outputs, including the adjustment direction, adjustment amplitude, and duration. When the tracking robot enters a region of abrupt curvature change, curvature response commands guide the robot to adjust in advance according to the direction determined by the balance results, enabling the robot to correct its attitude before the inertial trend causes a reversal. The curvature response commands also continuously correct the direction output as the trajectory change sensing sequence is updated in real time, forming a dynamic feedback loop that allows the tracking robot to maintain a stable trajectory under different curvature environments. Through this curvature response mechanism based on adaptive weight balance, the tracking robot can continuously adjust its direction of travel within curvature change regions, avoiding path reversals caused by the inertial trend, and maintaining a smooth trajectory transition and stable attitude.

[0074] Through the above implementation steps, an adaptive weight allocation relationship for curvature abrupt change nodes was established based on the trajectory inertial offset signal, achieving a dynamic balance between the inertial continuity trend and the real-time offset trend. A curvature response command was generated based on the balance result. This curvature response command operates in real-time during the robot's operation. By weakening the inertial continuity trend and strengthening the real-time offset trend, the robot can adjust its direction in advance during the initial stage of trajectory curvature changes, reducing the trajectory reversal delay effect and maintaining trajectory continuity and operational stability.

[0075] Step 4: Based on the curvature response command, the environmental recognition features contained in the trajectory image data are fused with the spatial positioning data between the sleepers to form a local attitude correction command stream. The trajectory of the track-following robot is continuously adjusted according to the local attitude correction command stream, so as to achieve a smooth transition of the trajectory turning segment and avoid the occurrence of direction reversal.

[0076] The specific steps for fusing environmental recognition features contained in the trajectory image data with spatial positioning data between sleepers to form a local attitude correction command stream are as follows:

[0077] After acquiring the curvature response command, the spatial region corresponding to the command is extracted from the trajectory image data, and environmental features are identified. The trajectory image data is continuously acquired by visual acquisition devices located in front of and below the track-following robot, containing a continuous sequence of images of sleepers, the track centerline, track edge lines, gaps between tracks, tunnel walls, trackside obstacles, and track support structures. Based on the spatial range specified by the curvature response command, the corresponding time segment in the trajectory image sequence is extracted, and the geometry, arrangement direction, sleeper spacing, spatial orientation of the track centerline, curvature trend of the track edge lines, and boundary features on both sides of the track are identified frame by frame. By analyzing the changes in sleeper arrangement, the directional shift of the track centerline, and the curvature changes of the edge lines in continuous image frames, the spatial structural characteristics and directional variation patterns of the track in the current area are determined. Further analysis of the lighting distribution, shadow projection, and depth relationship of trackside structures in the trajectory image data is conducted to identify the spatial hierarchy of the track and the relative position of the track edge with the surrounding environment. The aforementioned feature data are mapped to the time sequence of curvature response commands according to the time index, ensuring that the environmental identification features and curvature response commands are synchronized in time and space, providing a unified data reference for subsequent fusion with spatial positioning data.

[0078] After obtaining environmental recognition features from the trajectory image data, these features are fused with the spatial positioning data between sleepers based on the spatial position parameters of the curvature response command to form a comprehensive dataset reflecting the true spatial structure of the track. The spatial positioning data between sleepers is collected by the track-inspecting robot during its movement via a positioning device. This data includes the center point coordinates of each sleeper, the sleeper spacing, the sleeper's offset distance relative to the track centerline, the sleeper surface inclination angle, and the normal information of the plane containing the sleeper. Through spatial calibration using the curvature response command, this spatial positioning data is matched with the sleeper recognition results in the trajectory image data. During the matching process, firstly, using the track centerline as a reference, the spatial positioning data of the sleeper components is projected onto the coordinate space of the trajectory image data, establishing a correspondence between the actual spatial position of the sleeper and its shape in the image. Then, the spacing changes of each sleeper are compared with the sleeper arrangement direction identified in the image to correct spatial offset errors in the positioning data. Next, the normal angle of the sleeper surface is correlated with the spatial orientation of the track edge line to determine the lateral tilt characteristics and vertical attitude changes of the track within the current area. Through this series of operations, environmental recognition features and sleeper spatial positioning data are spatially fused to obtain a comprehensive trajectory spatial structure description that includes track geometry, sleeper distribution, and attitude inclination. This structural data reflects the actual spatial state of the track and is consistent with the directional parameters of the curvature response command, providing continuous input data for generating a local attitude correction command stream.

[0079] After fusing environmental recognition features from trajectory image data and spatial positioning data between sleepers, a local attitude correction command stream is generated by combining the directional information of curvature response commands. The curvature response commands include the target direction of trajectory adjustment, the range of curvature change, and the time interval for action. Based on the trajectory adjustment direction calibrated in the curvature response commands, the corresponding spatial path is identified in the fused comprehensive trajectory spatial structure data. Using sleeper spacing, track centerline direction, track edge line turning angle, and track tilt angle as parameters, the spatial direction change of the trajectory is calculated, and these changes are mapped to the attitude adjustment amount of the tracking robot. The generated local attitude correction command stream consists of continuous attitude adjustment commands, each including trajectory adjustment direction, attitude correction angle, adjustment magnitude, and adjustment duration. The trajectory adjustment direction determines the robot's direction of travel in space, the attitude correction angle controls the robot's attitude changes in the lateral and longitudinal directions, the adjustment magnitude limits the spatial range of each correction, and the adjustment duration ensures the continuous and smooth attitude change process. The local attitude correction command stream is output in chronological order, ensuring that the tracking robot can continuously perform attitude adjustments before and after curvature abrupt change nodes. The output of the command stream maintains a strict time correspondence with the curvature response command, enabling the tracking robot to continuously correct its attitude based on the latest curvature response information during changes in track curvature.

[0080] Finally, the trajectory of the tracking robot is continuously adjusted according to the generated local attitude correction command stream to achieve a smooth transition in the trajectory turning segment and avoid direction reversal. During implementation, the tracking robot executes each attitude adjustment command in the local attitude correction command stream sequentially. First, the robot chassis's forward direction is adjusted according to the trajectory adjustment direction parameter in the command, ensuring the direction of travel is consistent with the centerline of the track. Then, the lateral tilt angle and longitudinal pitch angle of the robot body are gradually adjusted according to the attitude correction angle, maintaining the robot's center of gravity balance and attitude stability during the turning process. Next, the rate of change of the turning angle is controlled according to the adjustment magnitude, ensuring a continuous transition in angle changes during the turning process and preventing sudden changes in trajectory direction from causing robot attitude swaying. Finally, the execution time of the attitude correction action is controlled according to the adjustment duration parameter, keeping the turning process synchronized with changes in track curvature. Throughout the entire trajectory adjustment process, the tracking robot continuously monitors changes in the track space structure based on the combined effect of curvature response commands and the local attitude correction command stream, adjusting its attitude in advance before entering the curvature abrupt change region, maintaining attitude stability during curvature changes, and naturally transitioning to a straight-line running state after the turning is completed. Through this continuous adjustment method, the track-following robot can achieve a smooth transition in the turning segment under different curvature environments, avoiding the problems of direction reversal and trajectory deviation caused by the continuation of inertia, and maintaining the continuity of the travel path and the stability of the attitude.

[0081] Through the above steps, the environmental recognition features contained in the trajectory image data and the spatial positioning data between sleepers are spatially fused according to the curvature response command, generating a complete local attitude correction command stream. This command stream is then used to continuously adjust the trajectory of the track-following robot, thereby achieving a smooth transition during trajectory turning segments and effectively preventing directional reversal. The entire process achieves unified fusion and dynamic linkage of visual environmental information, spatial positioning data, and curvature response information, enabling the track-following robot to maintain the continuity of trajectory movement and directional stability in complex track curvature environments.

[0082] Step 5: Based on the local attitude correction command stream, the visual dynamic distribution information in front of the trajectory is fused to generate the path inertia suppression output. Based on the path inertia suppression output, the travel path of the track-following robot is proactively adjusted in the curvature change region to ensure the travel continuity and attitude stability of the track-following robot in complex curves.

[0083] The specific steps for generating path inertia suppression output based on local attitude correction command stream and fusing visual dynamic distribution information ahead of the trajectory are as follows:

[0084] Based on the local attitude correction command stream generated in the previous stage, the visual dynamic distribution information in front of the trajectory is collected and identified to obtain the geometric change characteristics of the trajectory in space. During the operation of the track-following robot, the local attitude correction command stream outputs data such as trajectory adjustment direction, attitude correction angle, adjustment amplitude, and duration in real time. These data reflect the trajectory adjustment status of the robot at the current moment. According to the trajectory adjustment direction and range of action indicated in the local attitude correction command stream, the track-following robot continuously images the area in front of the trajectory through the visual acquisition device. The visual acquisition device collects a continuous image sequence along the track direction, including sleepers, track centerline, track edgeline, tunnel inner wall, and trackside obstacles. The image range covers a spatial area at a certain distance in front of the curvature abrupt change node. By analyzing the depth distribution, illumination changes, texture direction, and boundary morphology of the continuous image frames, the geometric change trend of the trajectory in space is identified, including the offset direction of the track centerline, changes in sleeper spacing, curvature of the track edgeline, and symmetry of the environmental structure on both sides of the track. To ensure consistency between visual information and attitude data, the temporal index of each visual image frame is synchronized with the temporal index of the local attitude correction command stream, ensuring that the visual dynamic distribution information remains consistent with the trajectory attitude adjustment state. The synchronized visual dynamic distribution information reflects the spatial changes ahead of the trajectory, providing spatial look-ahead data for the subsequent generation of path inertia suppression output.

[0085] After acquiring the visual dynamic distribution information ahead of the trajectory, it is fused with the local attitude correction command stream to generate a comprehensive dataset containing trajectory attitude and spatial geometric features. The local attitude correction command stream records the attitude change process of the tracking robot within the current curvature region, while the visual dynamic distribution information reflects the spatial change trend of the trajectory ahead. To achieve the fusion of the two types of data, firstly, based on the temporal correspondence between the two, the trajectory adjustment direction in the local attitude correction command stream is matched with the track centerline direction in the visual dynamic distribution information to ensure that the current attitude adjustment is consistent with the trend direction of the trajectory ahead. Then, the curvature of the track edge line in the visual dynamic distribution information is correlated with the attitude correction angle in the local attitude correction command stream to establish a spatial mapping relationship between track curvature changes and robot attitude adjustments. Finally, the change in sleeper spacing ahead of the trajectory is compared with the adjustment magnitude of the trajectory correction to determine the continuity of attitude changes in the trajectory extension direction. The fused dataset contains historical attitude adjustment information of the trajectory, spatial geometric features of the trajectory ahead, and the spatial correspondence between the two. This dataset can simultaneously reflect the robot's attitude state within the current curvature region and the spatial change pattern of the trajectory extension direction, providing a complete input basis for generating path inertia suppression output. During the data fusion process, all information is kept in time synchronization, enabling the path inertia suppression output to be dynamically adjusted based on the real-time trajectory status and the trend of the trajectory ahead.

[0086] After obtaining the fused trajectory attitude and visual distribution integrated data, a path inertia suppression output is generated, and based on this output, the robot's path is proactively adjusted within the curvature abrupt change region. The path inertia suppression output uses the fused data as input, integrating the trajectory direction information from the local attitude correction command stream with the curvature change characteristics of the forward visual dynamic distribution information to generate a continuous output containing the path adjustment direction, inertia suppression amplitude, and look-ahead adjustment duration. Specifically, firstly, the path adjustment direction is determined based on the curvature direction and rate of curvature change of the track centerline in the fused data, and this direction serves as the robot's correction direction within the curvature abrupt change region. Then, the inertia suppression amplitude is determined based on the attitude correction angle in the local attitude correction command stream and the degree of change in sleeper spacing in the visual distribution information, allowing the robot to reduce attitude deviation before the inertia continuation trend appears. Next, the look-ahead adjustment duration is determined based on the spatial location of the curvature abrupt change node and the extension direction of the track centerline, ensuring that the path inertia suppression output acts before the robot enters the curvature abrupt change region and gradually weakens after passing through it. The path inertia suppression output is sent in a continuous sequence. Each output value includes a direction vector, attitude adjustment amount, and duration, guiding the tracking robot to gradually adjust its attitude and path orientation as curvature changes. When the tracking robot approaches a curvature abrupt change region, the path inertia suppression output guides the robot to gradually correct its trajectory, maintaining attitude balance at curvature abrupt change nodes and smoothly resuming its straight-line direction when exiting a curve. Through this proactive adjustment method, the tracking robot can suppress the path deviation effect caused by trajectory inertia continuation in advance, avoiding attitude reversal and directional swing, and ensuring the continuity of trajectory turning and the stability of the travel process.

[0087] Through the above steps, the fusion of dynamic visual distribution information ahead of the trajectory is achieved based on the local attitude correction command stream, generating a path inertia suppression output. Based on this output, the robot's path is proactively adjusted. This process combines attitude correction information with the spatial distribution information of the trajectory ahead, enabling the robot to predict and respond to future trajectory changes within regions of abrupt changes in track curvature. It allows for path and attitude adjustments before curvature changes occur. The path inertia suppression output, by continuously applying to the robot's control commands, weakens the inertial continuation trend in real time, ensuring the robot maintains continuity of its direction of travel and stability of its attitude in complex curves, providing the final dynamic adjustment support for the entire positioning error compensation process.

[0088] This invention introduces a collaborative processing mechanism combining trajectory change sensing sequences and trajectory inertial offset signals, enabling a track-following robot to continuously perceive dynamic changes in the curvature of the track space during operation and continuously characterize the trajectory trend before and after curvature abrupt changes. By distinguishing between stable track segments and curvature abrupt change segments, and generating curvature response commands based on this distinction, the model's inertial dependence on existing trajectory trends is effectively reduced, allowing the positioning error compensation process to adjust in real time according to changes in track geometry. Therefore, the track-following robot can maintain the continuity of path judgment in complex track environments, reducing trajectory reversals and directional deviations caused by prediction lag, and improving the overall stability and reliability of positioning results.

[0089] Guided by curvature response commands, this invention integrates environmental recognition features, spatial positioning data between sleepers, and visual dynamic distribution information ahead of the track to construct a proactive adjustment process that flows from local attitude correction commands to path inertia suppression output. This process enables the track-inspecting robot to gradually correct its attitude and direction of travel before entering curves or areas of sudden curvature changes, thus achieving a smooth transition within turning sections. This proactive path adjustment method avoids problems such as direction reversal and discontinuous travel, enhancing the continuity of travel and attitude stability of the track-inspecting robot under complex curvature conditions, and providing a safer and more reliable operational guarantee for long-term autonomous inspection operations.

[0090] The foregoing has only described certain exemplary embodiments of the present invention by way of illustration. Undoubtedly, those skilled in the art can modify the described embodiments in various ways without departing from the spirit and scope of the present invention. Therefore, the foregoing drawings and descriptions are illustrative in nature and should not be construed as limiting the scope of protection of the claims of the present invention.

Claims

1. A method for compensating positioning errors of a track-following robot based on a large model, characterized in that, Includes the following steps: Step 1: Extract the spatial curvature features from the trajectory image data and inertial signal data collected during the operation of the track-following robot, and generate a trajectory change perception sequence based on the extracted spatial curvature features; Step 2: Based on the time window data consisting of multiple consecutive time points selected backward from the current timestamp in the trajectory change sensing sequence, the continuity characteristics of the stable track segment and the curvature change segment are dynamically separated. The historical running trend of the track is differentially mapped with the real-time trajectory offset data to obtain the trajectory inertial offset signal. Step 3: Based on the trajectory inertial offset signal, establish an adaptive weight allocation relationship for curvature change nodes, dynamically balance the inertial continuity trend and the real-time offset trend, and generate curvature response commands based on the balance results. Step 4: Based on the curvature response command, the environmental recognition features contained in the trajectory image data are fused with the spatial positioning data between the sleepers to form a local attitude correction command stream. The trajectory of the track-inspecting robot is continuously adjusted according to the local attitude correction command stream. Step 5: Based on the local attitude correction command stream, the visual dynamic distribution information in front of the trajectory is fused to generate path inertia suppression output. The travel path of the tracking robot is adjusted in the curvature change region according to the path inertia suppression output.

2. The method for compensating for positioning errors of a track-following robot based on a large model according to claim 1, characterized in that, The steps for generating a trajectory change sensing sequence based on the extracted spatial curvature features include: As the track-following robot moves along the track, it continuously acquires trajectory image data and simultaneously outputs inertial signal data that is time-aligned with the trajectory image data at the same time intervals. Based on the geometric structural features in the trajectory image data, the spatial morphology information of the track is extracted, and the angular velocity change and linear acceleration direction in the inertial signal data are matched with the curved area of ​​the track in the trajectory image data to generate a set of spatial curvature features that include the trajectory position, direction change and attitude curvature. Based on the set of spatial curvature features, with the direction of travel of the track-following robot as the main axis, the spatial curvature features corresponding to each moment are arranged in chronological order and a continuous correlation is established to generate a trajectory change perception sequence. Based on the curvature change relationship at each time point in the trajectory change sensing sequence, the rate of change of trajectory curvature is calculated. When the rate of change exceeds the set change threshold, the corresponding spatial position is determined as a curvature mutation node, and a mutation node chain is established.

3. The method for compensating for positioning errors of a track-following robot based on a large model according to claim 2, characterized in that, The steps for obtaining the trajectory inertial offset signal by differential mapping between the historical trajectory trend and real-time trajectory offset data include: During track operation, the track-following robot reads the trajectory space curvature value, attitude angle, direction vector and track displacement information frame by frame from the time window data interval consisting of multiple time points continuously selected forward from the current timestamp in the trajectory change perception sequence. It also performs spatial reconstruction by combining the sleeper, track edge line and center line features in the trajectory image data. At the same time, it matches the angular velocity change, linear acceleration direction and attitude angle change amplitude in the inertial signal data with the trajectory curvature change to build a structured trajectory operation dataset. The continuous feature separation is performed on the time series data of the time window data interval composed of multiple time points continuously selected forward from the current timestamp in the trajectory change perception sequence. The part of the trajectory space curvature value change amplitude continuously less than the amplitude threshold is classified into the stable segment set, and the part of the trajectory curvature change rate exceeds the change threshold is classified into the abrupt segment set. Smooth transition processing is performed in the boundary area between the track stable segment and the curvature abrupt segment to form the trajectory segment structure. Based on the feature data in the stable segment set and the abrupt segment set, differential mapping is performed on the trajectory direction change, attitude deflection and orbit centerline displacement difference according to the time correspondence to generate the trajectory inertial offset signal.

4. The method for compensating for positioning errors of a tracking robot based on a large model according to claim 3, characterized in that, During the generation of the trajectory inertial offset signal, the differential mapping of the trajectory direction change, attitude deflection and orbit centerline displacement difference is continuously calculated in time order, and the three parameters of inertial continuation direction, offset amplitude and duration are recorded in the trajectory inertial offset signal respectively.

5. The method for compensating for positioning errors of a track-following robot based on a large model according to claim 4, characterized in that, The steps for generating curvature response instructions based on the equilibrium results include: Based on the trajectory inertial offset signal, spatial location identification and temporal correlation are performed on the curvature change nodes. The trajectory inertial continuation direction, offset amplitude, duration and attitude angle change are extracted and matched with the sleeper spacing change, track edge line turning angle and track center line bending radius in the trajectory image data to construct a node comprehensive dataset containing inertial offset features and curvature change features. Based on the node comprehensive dataset, an adaptive weight allocation relationship is established for each curvature change node. According to the correspondence between the trajectory inertial offset signal and the real-time offset trend in the two dimensions of time and space, the ratio of the inertial continuation trend weight to the real-time offset trend weight is adjusted to generate a dynamically balanced weight allocation result. Based on the weight allocation results and trajectory inertial offset signal, combined with the spatial geometric features and attitude change features of curvature change nodes, the adjustment direction, adjustment amplitude and action time of curvature response commands are determined, and curvature response commands are generated in a continuous output manner.

6. The method for compensating for positioning errors of a tracking robot based on a large model according to claim 5, characterized in that, The adjustment direction of the curvature response command is determined by the vector superposition of the inertial continuation trend direction after dynamic equilibrium and the real-time offset trend direction. The adjustment range is set according to the offset range of the trajectory inertial offset signal, and the action time is determined according to the duration of the trajectory inertial offset signal. The curvature response command is output before the track-following robot enters the curvature change region.

7. The method for compensating for positioning errors of a track-following robot based on a large model according to claim 5, characterized in that, The steps of fusing environmental recognition features contained in trajectory image data with spatial positioning data between sleepers to form a local attitude correction command stream include: After obtaining the curvature response command, the spatial region corresponding to the curvature response command is extracted from the trajectory image data, and the environmental recognition features are synchronized with the curvature response command in time and space according to the time index. Based on the spatial position parameters of the curvature response command, the environmental identification features are fused with the spatial positioning data between sleepers. By matching the coordinates of the track centerline, comparing the sleeper spacing, and associating the normal angle of the sleeper surface, a comprehensive trajectory spatial structure description is formed, which includes the track geometry, sleeper distribution pattern, and attitude inclination angle. Combined with the adjustment direction of the curvature response command, the trajectory adjustment direction is identified in the fused integrated trajectory space structure. Based on the sleeper spacing, track centerline direction, track edge line turning angle and track tilt angle, a continuous local attitude correction command flow is generated. The local attitude correction command flow includes trajectory adjustment direction, attitude correction angle, adjustment magnitude and adjustment duration. The trajectory of the tracking robot is continuously adjusted based on the generated local attitude correction command stream.

8. The method for compensating for positioning errors of a tracking robot based on a large model according to claim 7, characterized in that, During the continuous adjustment of the trajectory of the track-following robot based on the generated local attitude correction command stream, the trajectory adjustment includes continuous correction of the forward direction within the range of curvature response command, and real-time update of attitude correction angle when the sleeper spacing changes synchronously with the turning angle of the track edge line.

9. The method for compensating for positioning errors of a track-following robot based on a large model according to claim 7, characterized in that, The steps for generating path inertia suppression output based on local attitude correction command stream and fusing visual dynamic distribution information ahead of the trajectory include: Based on the local attitude correction command stream, the visual dynamic distribution information in front of the track is collected and identified. According to the trajectory adjustment direction and range of action in the local attitude correction command stream, the visual acquisition device acquires an image sequence containing sleepers, track center line, track edge line, tunnel inner wall and trackside obstacles. The visual dynamic distribution information is synchronized with the attitude data according to the time index to obtain the spatial geometric features in front of the track. After acquiring the visual dynamic distribution information in front of the trajectory, it is fused with the local attitude correction command stream. Based on the time correspondence between the visual dynamic distribution information and the local attitude correction command stream, the trajectory adjustment direction is matched with the direction of the orbit centerline to form a comprehensive dataset containing trajectory attitude and spatial geometric features. Based on the curvature direction, curvature change rate, and attitude correction angle of the track centerline, a path inertia suppression output is generated, which includes the path adjustment direction, inertia suppression amplitude, and adjustment duration. The travel path of the track-following robot is adjusted in the curvature change region by continuous output.

10. The method for compensating for positioning errors of a track-following robot based on a large model according to claim 9, characterized in that, During the generation process, the path inertia suppression output determines the path adjustment direction based on the curvature direction and rate of curvature change of the track centerline, and determines the inertia suppression amplitude based on the attitude correction angle in the local attitude correction command stream and the sleeper spacing change in the visual dynamic distribution information.

Images